1 /* 2 * Copyright (c) 1998, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/assembler.inline.hpp" 27 #include "code/compiledIC.hpp" 28 #include "code/debugInfo.hpp" 29 #include "code/debugInfoRec.hpp" 30 #include "compiler/compileBroker.hpp" 31 #include "compiler/oopMap.hpp" 32 #include "memory/allocation.inline.hpp" 33 #include "opto/callnode.hpp" 34 #include "opto/cfgnode.hpp" 35 #include "opto/locknode.hpp" 36 #include "opto/machnode.hpp" 37 #include "opto/output.hpp" 38 #include "opto/regalloc.hpp" 39 #include "opto/runtime.hpp" 40 #include "opto/subnode.hpp" 41 #include "opto/type.hpp" 42 #include "runtime/handles.inline.hpp" 43 #include "utilities/xmlstream.hpp" 44 45 #ifndef PRODUCT 46 #define DEBUG_ARG(x) , x 47 #else 48 #define DEBUG_ARG(x) 49 #endif 50 51 // Convert Nodes to instruction bits and pass off to the VM 52 void Compile::Output() { 53 // RootNode goes 54 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" ); 55 56 // The number of new nodes (mostly MachNop) is proportional to 57 // the number of java calls and inner loops which are aligned. 58 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 59 C->inner_loops()*(OptoLoopAlignment-1)), 60 "out of nodes before code generation" ) ) { 61 return; 62 } 63 // Make sure I can find the Start Node 64 Block *entry = _cfg->get_block(1); 65 Block *broot = _cfg->get_root_block(); 66 67 const StartNode *start = entry->head()->as_Start(); 68 69 // Replace StartNode with prolog 70 MachPrologNode *prolog = new (this) MachPrologNode(); 71 entry->map_node(prolog, 0); 72 _cfg->map_node_to_block(prolog, entry); 73 _cfg->unmap_node_from_block(start); // start is no longer in any block 74 75 // Virtual methods need an unverified entry point 76 77 if( is_osr_compilation() ) { 78 if( PoisonOSREntry ) { 79 // TODO: Should use a ShouldNotReachHereNode... 80 _cfg->insert( broot, 0, new (this) MachBreakpointNode() ); 81 } 82 } else { 83 if( _method && !_method->flags().is_static() ) { 84 // Insert unvalidated entry point 85 _cfg->insert( broot, 0, new (this) MachUEPNode() ); 86 } 87 88 } 89 90 91 // Break before main entry point 92 if( (_method && _method->break_at_execute()) 93 #ifndef PRODUCT 94 ||(OptoBreakpoint && is_method_compilation()) 95 ||(OptoBreakpointOSR && is_osr_compilation()) 96 ||(OptoBreakpointC2R && !_method) 97 #endif 98 ) { 99 // checking for _method means that OptoBreakpoint does not apply to 100 // runtime stubs or frame converters 101 _cfg->insert( entry, 1, new (this) MachBreakpointNode() ); 102 } 103 104 // Insert epilogs before every return 105 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 106 Block* block = _cfg->get_block(i); 107 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point? 108 Node* m = block->end(); 109 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) { 110 MachEpilogNode* epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 111 block->add_inst(epilog); 112 _cfg->map_node_to_block(epilog, block); 113 } 114 } 115 } 116 117 # ifdef ENABLE_ZAP_DEAD_LOCALS 118 if (ZapDeadCompiledLocals) { 119 Insert_zap_nodes(); 120 } 121 # endif 122 123 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1); 124 blk_starts[0] = 0; 125 126 // Initialize code buffer and process short branches. 127 CodeBuffer* cb = init_buffer(blk_starts); 128 129 if (cb == NULL || failing()) { 130 return; 131 } 132 133 ScheduleAndBundle(); 134 135 #ifndef PRODUCT 136 if (trace_opto_output()) { 137 tty->print("\n---- After ScheduleAndBundle ----\n"); 138 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 139 tty->print("\nBB#%03d:\n", i); 140 Block* block = _cfg->get_block(i); 141 for (uint j = 0; j < block->number_of_nodes(); j++) { 142 Node* n = block->get_node(j); 143 OptoReg::Name reg = _regalloc->get_reg_first(n); 144 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 145 n->dump(); 146 } 147 } 148 } 149 #endif 150 151 if (failing()) { 152 return; 153 } 154 155 BuildOopMaps(); 156 157 if (failing()) { 158 return; 159 } 160 161 fill_buffer(cb, blk_starts); 162 } 163 164 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 165 // Determine if we need to generate a stack overflow check. 166 // Do it if the method is not a stub function and 167 // has java calls or has frame size > vm_page_size/8. 168 return (UseStackBanging && stub_function() == NULL && 169 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3)); 170 } 171 172 bool Compile::need_register_stack_bang() const { 173 // Determine if we need to generate a register stack overflow check. 174 // This is only used on architectures which have split register 175 // and memory stacks (ie. IA64). 176 // Bang if the method is not a stub function and has java calls 177 return (stub_function() == NULL && has_java_calls()); 178 } 179 180 # ifdef ENABLE_ZAP_DEAD_LOCALS 181 182 183 // In order to catch compiler oop-map bugs, we have implemented 184 // a debugging mode called ZapDeadCompilerLocals. 185 // This mode causes the compiler to insert a call to a runtime routine, 186 // "zap_dead_locals", right before each place in compiled code 187 // that could potentially be a gc-point (i.e., a safepoint or oop map point). 188 // The runtime routine checks that locations mapped as oops are really 189 // oops, that locations mapped as values do not look like oops, 190 // and that locations mapped as dead are not used later 191 // (by zapping them to an invalid address). 192 193 int Compile::_CompiledZap_count = 0; 194 195 void Compile::Insert_zap_nodes() { 196 bool skip = false; 197 198 199 // Dink with static counts because code code without the extra 200 // runtime calls is MUCH faster for debugging purposes 201 202 if ( CompileZapFirst == 0 ) ; // nothing special 203 else if ( CompileZapFirst > CompiledZap_count() ) skip = true; 204 else if ( CompileZapFirst == CompiledZap_count() ) 205 warning("starting zap compilation after skipping"); 206 207 if ( CompileZapLast == -1 ) ; // nothing special 208 else if ( CompileZapLast < CompiledZap_count() ) skip = true; 209 else if ( CompileZapLast == CompiledZap_count() ) 210 warning("about to compile last zap"); 211 212 ++_CompiledZap_count; // counts skipped zaps, too 213 214 if ( skip ) return; 215 216 217 if ( _method == NULL ) 218 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care 219 220 // Insert call to zap runtime stub before every node with an oop map 221 for( uint i=0; i<_cfg->number_of_blocks(); i++ ) { 222 Block *b = _cfg->get_block(i); 223 for ( uint j = 0; j < b->number_of_nodes(); ++j ) { 224 Node *n = b->get_node(j); 225 226 // Determining if we should insert a zap-a-lot node in output. 227 // We do that for all nodes that has oopmap info, except for calls 228 // to allocation. Calls to allocation passes in the old top-of-eden pointer 229 // and expect the C code to reset it. Hence, there can be no safepoints between 230 // the inlined-allocation and the call to new_Java, etc. 231 // We also cannot zap monitor calls, as they must hold the microlock 232 // during the call to Zap, which also wants to grab the microlock. 233 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL); 234 if ( insert ) { // it is MachSafePoint 235 if ( !n->is_MachCall() ) { 236 insert = false; 237 } else if ( n->is_MachCall() ) { 238 MachCallNode* call = n->as_MachCall(); 239 if (call->entry_point() == OptoRuntime::new_instance_Java() || 240 call->entry_point() == OptoRuntime::new_array_Java() || 241 call->entry_point() == OptoRuntime::multianewarray2_Java() || 242 call->entry_point() == OptoRuntime::multianewarray3_Java() || 243 call->entry_point() == OptoRuntime::multianewarray4_Java() || 244 call->entry_point() == OptoRuntime::multianewarray5_Java() || 245 call->entry_point() == OptoRuntime::slow_arraycopy_Java() || 246 call->entry_point() == OptoRuntime::complete_monitor_locking_Java() 247 ) { 248 insert = false; 249 } 250 } 251 if (insert) { 252 Node *zap = call_zap_node(n->as_MachSafePoint(), i); 253 b->insert_node(zap, j); 254 _cfg->map_node_to_block(zap, b); 255 ++j; 256 } 257 } 258 } 259 } 260 } 261 262 263 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) { 264 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type(); 265 CallStaticJavaNode* ideal_node = 266 new (this) CallStaticJavaNode( tf, 267 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()), 268 "call zap dead locals stub", 0, TypePtr::BOTTOM); 269 // We need to copy the OopMap from the site we're zapping at. 270 // We have to make a copy, because the zap site might not be 271 // a call site, and zap_dead is a call site. 272 OopMap* clone = node_to_check->oop_map()->deep_copy(); 273 274 // Add the cloned OopMap to the zap node 275 ideal_node->set_oop_map(clone); 276 return _matcher->match_sfpt(ideal_node); 277 } 278 279 bool Compile::is_node_getting_a_safepoint( Node* n) { 280 // This code duplicates the logic prior to the call of add_safepoint 281 // below in this file. 282 if( n->is_MachSafePoint() ) return true; 283 return false; 284 } 285 286 # endif // ENABLE_ZAP_DEAD_LOCALS 287 288 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 289 // of a loop. When aligning a loop we need to provide enough instructions 290 // in cpu's fetch buffer to feed decoders. The loop alignment could be 291 // avoided if we have enough instructions in fetch buffer at the head of a loop. 292 // By default, the size is set to 999999 by Block's constructor so that 293 // a loop will be aligned if the size is not reset here. 294 // 295 // Note: Mach instructions could contain several HW instructions 296 // so the size is estimated only. 297 // 298 void Compile::compute_loop_first_inst_sizes() { 299 // The next condition is used to gate the loop alignment optimization. 300 // Don't aligned a loop if there are enough instructions at the head of a loop 301 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 302 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 303 // equal to 11 bytes which is the largest address NOP instruction. 304 if (MaxLoopPad < OptoLoopAlignment - 1) { 305 uint last_block = _cfg->number_of_blocks() - 1; 306 for (uint i = 1; i <= last_block; i++) { 307 Block* block = _cfg->get_block(i); 308 // Check the first loop's block which requires an alignment. 309 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) { 310 uint sum_size = 0; 311 uint inst_cnt = NumberOfLoopInstrToAlign; 312 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 313 314 // Check subsequent fallthrough blocks if the loop's first 315 // block(s) does not have enough instructions. 316 Block *nb = block; 317 while(inst_cnt > 0 && 318 i < last_block && 319 !_cfg->get_block(i + 1)->has_loop_alignment() && 320 !nb->has_successor(block)) { 321 i++; 322 nb = _cfg->get_block(i); 323 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 324 } // while( inst_cnt > 0 && i < last_block ) 325 326 block->set_first_inst_size(sum_size); 327 } // f( b->head()->is_Loop() ) 328 } // for( i <= last_block ) 329 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 330 } 331 332 // The architecture description provides short branch variants for some long 333 // branch instructions. Replace eligible long branches with short branches. 334 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) { 335 // Compute size of each block, method size, and relocation information size 336 uint nblocks = _cfg->number_of_blocks(); 337 338 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 339 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 340 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 341 342 // Collect worst case block paddings 343 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks); 344 memset(block_worst_case_pad, 0, nblocks * sizeof(int)); 345 346 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 347 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 348 349 bool has_short_branch_candidate = false; 350 351 // Initialize the sizes to 0 352 code_size = 0; // Size in bytes of generated code 353 stub_size = 0; // Size in bytes of all stub entries 354 // Size in bytes of all relocation entries, including those in local stubs. 355 // Start with 2-bytes of reloc info for the unvalidated entry point 356 reloc_size = 1; // Number of relocation entries 357 358 // Make three passes. The first computes pessimistic blk_starts, 359 // relative jmp_offset and reloc_size information. The second performs 360 // short branch substitution using the pessimistic sizing. The 361 // third inserts nops where needed. 362 363 // Step one, perform a pessimistic sizing pass. 364 uint last_call_adr = max_uint; 365 uint last_avoid_back_to_back_adr = max_uint; 366 uint nop_size = (new (this) MachNopNode())->size(_regalloc); 367 for (uint i = 0; i < nblocks; i++) { // For all blocks 368 Block* block = _cfg->get_block(i); 369 370 // During short branch replacement, we store the relative (to blk_starts) 371 // offset of jump in jmp_offset, rather than the absolute offset of jump. 372 // This is so that we do not need to recompute sizes of all nodes when 373 // we compute correct blk_starts in our next sizing pass. 374 jmp_offset[i] = 0; 375 jmp_size[i] = 0; 376 jmp_nidx[i] = -1; 377 DEBUG_ONLY( jmp_target[i] = 0; ) 378 DEBUG_ONLY( jmp_rule[i] = 0; ) 379 380 // Sum all instruction sizes to compute block size 381 uint last_inst = block->number_of_nodes(); 382 uint blk_size = 0; 383 for (uint j = 0; j < last_inst; j++) { 384 Node* nj = block->get_node(j); 385 // Handle machine instruction nodes 386 if (nj->is_Mach()) { 387 MachNode *mach = nj->as_Mach(); 388 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 389 reloc_size += mach->reloc(); 390 if (mach->is_MachCall()) { 391 // add size information for trampoline stub 392 // class CallStubImpl is platform-specific and defined in the *.ad files. 393 stub_size += CallStubImpl::size_call_trampoline(); 394 reloc_size += CallStubImpl::reloc_call_trampoline(); 395 396 MachCallNode *mcall = mach->as_MachCall(); 397 // This destination address is NOT PC-relative 398 399 mcall->method_set((intptr_t)mcall->entry_point()); 400 401 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 402 stub_size += CompiledStaticCall::to_interp_stub_size(); 403 reloc_size += CompiledStaticCall::reloc_to_interp_stub(); 404 } 405 } else if (mach->is_MachSafePoint()) { 406 // If call/safepoint are adjacent, account for possible 407 // nop to disambiguate the two safepoints. 408 // ScheduleAndBundle() can rearrange nodes in a block, 409 // check for all offsets inside this block. 410 if (last_call_adr >= blk_starts[i]) { 411 blk_size += nop_size; 412 } 413 } 414 if (mach->avoid_back_to_back()) { 415 // Nop is inserted between "avoid back to back" instructions. 416 // ScheduleAndBundle() can rearrange nodes in a block, 417 // check for all offsets inside this block. 418 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 419 blk_size += nop_size; 420 } 421 } 422 if (mach->may_be_short_branch()) { 423 if (!nj->is_MachBranch()) { 424 #ifndef PRODUCT 425 nj->dump(3); 426 #endif 427 Unimplemented(); 428 } 429 assert(jmp_nidx[i] == -1, "block should have only one branch"); 430 jmp_offset[i] = blk_size; 431 jmp_size[i] = nj->size(_regalloc); 432 jmp_nidx[i] = j; 433 has_short_branch_candidate = true; 434 } 435 } 436 blk_size += nj->size(_regalloc); 437 // Remember end of call offset 438 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 439 last_call_adr = blk_starts[i]+blk_size; 440 } 441 // Remember end of avoid_back_to_back offset 442 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) { 443 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 444 } 445 } 446 447 // When the next block starts a loop, we may insert pad NOP 448 // instructions. Since we cannot know our future alignment, 449 // assume the worst. 450 if (i < nblocks - 1) { 451 Block* nb = _cfg->get_block(i + 1); 452 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 453 if (max_loop_pad > 0) { 454 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 455 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 456 // If either is the last instruction in this block, bump by 457 // max_loop_pad in lock-step with blk_size, so sizing 458 // calculations in subsequent blocks still can conservatively 459 // detect that it may the last instruction in this block. 460 if (last_call_adr == blk_starts[i]+blk_size) { 461 last_call_adr += max_loop_pad; 462 } 463 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 464 last_avoid_back_to_back_adr += max_loop_pad; 465 } 466 blk_size += max_loop_pad; 467 block_worst_case_pad[i + 1] = max_loop_pad; 468 } 469 } 470 471 // Save block size; update total method size 472 blk_starts[i+1] = blk_starts[i]+blk_size; 473 } 474 475 // Step two, replace eligible long jumps. 476 bool progress = true; 477 uint last_may_be_short_branch_adr = max_uint; 478 while (has_short_branch_candidate && progress) { 479 progress = false; 480 has_short_branch_candidate = false; 481 int adjust_block_start = 0; 482 for (uint i = 0; i < nblocks; i++) { 483 Block* block = _cfg->get_block(i); 484 int idx = jmp_nidx[i]; 485 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach(); 486 if (mach != NULL && mach->may_be_short_branch()) { 487 #ifdef ASSERT 488 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 489 int j; 490 // Find the branch; ignore trailing NOPs. 491 for (j = block->number_of_nodes()-1; j>=0; j--) { 492 Node* n = block->get_node(j); 493 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 494 break; 495 } 496 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity"); 497 #endif 498 int br_size = jmp_size[i]; 499 int br_offs = blk_starts[i] + jmp_offset[i]; 500 501 // This requires the TRUE branch target be in succs[0] 502 uint bnum = block->non_connector_successor(0)->_pre_order; 503 int offset = blk_starts[bnum] - br_offs; 504 if (bnum > i) { // adjust following block's offset 505 offset -= adjust_block_start; 506 } 507 508 // This block can be a loop header, account for the padding 509 // in the previous block. 510 int block_padding = block_worst_case_pad[i]; 511 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top"); 512 // In the following code a nop could be inserted before 513 // the branch which will increase the backward distance. 514 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr); 515 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block"); 516 517 if (needs_padding && offset <= 0) 518 offset -= nop_size; 519 520 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 521 // We've got a winner. Replace this branch. 522 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 523 524 // Update the jmp_size. 525 int new_size = replacement->size(_regalloc); 526 int diff = br_size - new_size; 527 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 528 // Conservatively take into accound padding between 529 // avoid_back_to_back branches. Previous branch could be 530 // converted into avoid_back_to_back branch during next 531 // rounds. 532 if (needs_padding && replacement->avoid_back_to_back()) { 533 jmp_offset[i] += nop_size; 534 diff -= nop_size; 535 } 536 adjust_block_start += diff; 537 block->map_node(replacement, idx); 538 mach->subsume_by(replacement, C); 539 mach = replacement; 540 progress = true; 541 542 jmp_size[i] = new_size; 543 DEBUG_ONLY( jmp_target[i] = bnum; ); 544 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 545 } else { 546 // The jump distance is not short, try again during next iteration. 547 has_short_branch_candidate = true; 548 } 549 } // (mach->may_be_short_branch()) 550 if (mach != NULL && (mach->may_be_short_branch() || 551 mach->avoid_back_to_back())) { 552 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 553 } 554 blk_starts[i+1] -= adjust_block_start; 555 } 556 } 557 558 #ifdef ASSERT 559 for (uint i = 0; i < nblocks; i++) { // For all blocks 560 if (jmp_target[i] != 0) { 561 int br_size = jmp_size[i]; 562 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 563 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 564 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]); 565 } 566 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 567 } 568 } 569 #endif 570 571 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 572 // after ScheduleAndBundle(). 573 574 // ------------------ 575 // Compute size for code buffer 576 code_size = blk_starts[nblocks]; 577 578 // Relocation records 579 reloc_size += 1; // Relo entry for exception handler 580 581 // Adjust reloc_size to number of record of relocation info 582 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 583 // a relocation index. 584 // The CodeBuffer will expand the locs array if this estimate is too low. 585 reloc_size *= 10 / sizeof(relocInfo); 586 } 587 588 //------------------------------FillLocArray----------------------------------- 589 // Create a bit of debug info and append it to the array. The mapping is from 590 // Java local or expression stack to constant, register or stack-slot. For 591 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 592 // entry has been taken care of and caller should skip it). 593 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 594 // This should never have accepted Bad before 595 assert(OptoReg::is_valid(regnum), "location must be valid"); 596 return (OptoReg::is_reg(regnum)) 597 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 598 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 599 } 600 601 602 ObjectValue* 603 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 604 for (int i = 0; i < objs->length(); i++) { 605 assert(objs->at(i)->is_object(), "corrupt object cache"); 606 ObjectValue* sv = (ObjectValue*) objs->at(i); 607 if (sv->id() == id) { 608 return sv; 609 } 610 } 611 // Otherwise.. 612 return NULL; 613 } 614 615 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 616 ObjectValue* sv ) { 617 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 618 objs->append(sv); 619 } 620 621 622 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 623 GrowableArray<ScopeValue*> *array, 624 GrowableArray<ScopeValue*> *objs ) { 625 assert( local, "use _top instead of null" ); 626 if (array->length() != idx) { 627 assert(array->length() == idx + 1, "Unexpected array count"); 628 // Old functionality: 629 // return 630 // New functionality: 631 // Assert if the local is not top. In product mode let the new node 632 // override the old entry. 633 assert(local == top(), "LocArray collision"); 634 if (local == top()) { 635 return; 636 } 637 array->pop(); 638 } 639 const Type *t = local->bottom_type(); 640 641 // Is it a safepoint scalar object node? 642 if (local->is_SafePointScalarObject()) { 643 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 644 645 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 646 if (sv == NULL) { 647 ciKlass* cik = t->is_oopptr()->klass(); 648 assert(cik->is_instance_klass() || 649 cik->is_array_klass(), "Not supported allocation."); 650 sv = new ObjectValue(spobj->_idx, 651 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 652 Compile::set_sv_for_object_node(objs, sv); 653 654 uint first_ind = spobj->first_index(sfpt->jvms()); 655 for (uint i = 0; i < spobj->n_fields(); i++) { 656 Node* fld_node = sfpt->in(first_ind+i); 657 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 658 } 659 } 660 array->append(sv); 661 return; 662 } 663 664 // Grab the register number for the local 665 OptoReg::Name regnum = _regalloc->get_reg_first(local); 666 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 667 // Record the double as two float registers. 668 // The register mask for such a value always specifies two adjacent 669 // float registers, with the lower register number even. 670 // Normally, the allocation of high and low words to these registers 671 // is irrelevant, because nearly all operations on register pairs 672 // (e.g., StoreD) treat them as a single unit. 673 // Here, we assume in addition that the words in these two registers 674 // stored "naturally" (by operations like StoreD and double stores 675 // within the interpreter) such that the lower-numbered register 676 // is written to the lower memory address. This may seem like 677 // a machine dependency, but it is not--it is a requirement on 678 // the author of the <arch>.ad file to ensure that, for every 679 // even/odd double-register pair to which a double may be allocated, 680 // the word in the even single-register is stored to the first 681 // memory word. (Note that register numbers are completely 682 // arbitrary, and are not tied to any machine-level encodings.) 683 #ifdef _LP64 684 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 685 array->append(new ConstantIntValue(0)); 686 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 687 } else if ( t->base() == Type::Long ) { 688 array->append(new ConstantIntValue(0)); 689 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 690 } else if ( t->base() == Type::RawPtr ) { 691 // jsr/ret return address which must be restored into a the full 692 // width 64-bit stack slot. 693 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 694 } 695 #else //_LP64 696 #ifdef SPARC 697 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 698 // For SPARC we have to swap high and low words for 699 // long values stored in a single-register (g0-g7). 700 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 701 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 702 } else 703 #endif //SPARC 704 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 705 // Repack the double/long as two jints. 706 // The convention the interpreter uses is that the second local 707 // holds the first raw word of the native double representation. 708 // This is actually reasonable, since locals and stack arrays 709 // grow downwards in all implementations. 710 // (If, on some machine, the interpreter's Java locals or stack 711 // were to grow upwards, the embedded doubles would be word-swapped.) 712 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 713 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 714 } 715 #endif //_LP64 716 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 717 OptoReg::is_reg(regnum) ) { 718 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 719 ? Location::float_in_dbl : Location::normal )); 720 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 721 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 722 ? Location::int_in_long : Location::normal )); 723 } else if( t->base() == Type::NarrowOop ) { 724 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 725 } else { 726 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 727 } 728 return; 729 } 730 731 // No register. It must be constant data. 732 switch (t->base()) { 733 case Type::Half: // Second half of a double 734 ShouldNotReachHere(); // Caller should skip 2nd halves 735 break; 736 case Type::AnyPtr: 737 array->append(new ConstantOopWriteValue(NULL)); 738 break; 739 case Type::AryPtr: 740 case Type::InstPtr: // fall through 741 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 742 break; 743 case Type::NarrowOop: 744 if (t == TypeNarrowOop::NULL_PTR) { 745 array->append(new ConstantOopWriteValue(NULL)); 746 } else { 747 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 748 } 749 break; 750 case Type::Int: 751 array->append(new ConstantIntValue(t->is_int()->get_con())); 752 break; 753 case Type::RawPtr: 754 // A return address (T_ADDRESS). 755 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 756 #ifdef _LP64 757 // Must be restored to the full-width 64-bit stack slot. 758 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 759 #else 760 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 761 #endif 762 break; 763 case Type::FloatCon: { 764 float f = t->is_float_constant()->getf(); 765 array->append(new ConstantIntValue(jint_cast(f))); 766 break; 767 } 768 case Type::DoubleCon: { 769 jdouble d = t->is_double_constant()->getd(); 770 #ifdef _LP64 771 array->append(new ConstantIntValue(0)); 772 array->append(new ConstantDoubleValue(d)); 773 #else 774 // Repack the double as two jints. 775 // The convention the interpreter uses is that the second local 776 // holds the first raw word of the native double representation. 777 // This is actually reasonable, since locals and stack arrays 778 // grow downwards in all implementations. 779 // (If, on some machine, the interpreter's Java locals or stack 780 // were to grow upwards, the embedded doubles would be word-swapped.) 781 jint *dp = (jint*)&d; 782 array->append(new ConstantIntValue(dp[1])); 783 array->append(new ConstantIntValue(dp[0])); 784 #endif 785 break; 786 } 787 case Type::Long: { 788 jlong d = t->is_long()->get_con(); 789 #ifdef _LP64 790 array->append(new ConstantIntValue(0)); 791 array->append(new ConstantLongValue(d)); 792 #else 793 // Repack the long as two jints. 794 // The convention the interpreter uses is that the second local 795 // holds the first raw word of the native double representation. 796 // This is actually reasonable, since locals and stack arrays 797 // grow downwards in all implementations. 798 // (If, on some machine, the interpreter's Java locals or stack 799 // were to grow upwards, the embedded doubles would be word-swapped.) 800 jint *dp = (jint*)&d; 801 array->append(new ConstantIntValue(dp[1])); 802 array->append(new ConstantIntValue(dp[0])); 803 #endif 804 break; 805 } 806 case Type::Top: // Add an illegal value here 807 array->append(new LocationValue(Location())); 808 break; 809 default: 810 ShouldNotReachHere(); 811 break; 812 } 813 } 814 815 // Determine if this node starts a bundle 816 bool Compile::starts_bundle(const Node *n) const { 817 return (_node_bundling_limit > n->_idx && 818 _node_bundling_base[n->_idx].starts_bundle()); 819 } 820 821 //--------------------------Process_OopMap_Node-------------------------------- 822 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 823 824 // Handle special safepoint nodes for synchronization 825 MachSafePointNode *sfn = mach->as_MachSafePoint(); 826 MachCallNode *mcall; 827 828 #ifdef ENABLE_ZAP_DEAD_LOCALS 829 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative"); 830 #endif 831 832 int safepoint_pc_offset = current_offset; 833 bool is_method_handle_invoke = false; 834 bool return_oop = false; 835 836 // Add the safepoint in the DebugInfoRecorder 837 if( !mach->is_MachCall() ) { 838 mcall = NULL; 839 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 840 } else { 841 mcall = mach->as_MachCall(); 842 843 // Is the call a MethodHandle call? 844 if (mcall->is_MachCallJava()) { 845 if (mcall->as_MachCallJava()->_method_handle_invoke) { 846 assert(has_method_handle_invokes(), "must have been set during call generation"); 847 is_method_handle_invoke = true; 848 } 849 } 850 851 // Check if a call returns an object. 852 if (mcall->return_value_is_used() && 853 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) { 854 return_oop = true; 855 } 856 safepoint_pc_offset += mcall->ret_addr_offset(); 857 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 858 } 859 860 // Loop over the JVMState list to add scope information 861 // Do not skip safepoints with a NULL method, they need monitor info 862 JVMState* youngest_jvms = sfn->jvms(); 863 int max_depth = youngest_jvms->depth(); 864 865 // Allocate the object pool for scalar-replaced objects -- the map from 866 // small-integer keys (which can be recorded in the local and ostack 867 // arrays) to descriptions of the object state. 868 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 869 870 // Visit scopes from oldest to youngest. 871 for (int depth = 1; depth <= max_depth; depth++) { 872 JVMState* jvms = youngest_jvms->of_depth(depth); 873 int idx; 874 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 875 // Safepoints that do not have method() set only provide oop-map and monitor info 876 // to support GC; these do not support deoptimization. 877 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 878 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 879 int num_mon = jvms->nof_monitors(); 880 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 881 "JVMS local count must match that of the method"); 882 883 // Add Local and Expression Stack Information 884 885 // Insert locals into the locarray 886 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 887 for( idx = 0; idx < num_locs; idx++ ) { 888 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 889 } 890 891 // Insert expression stack entries into the exparray 892 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 893 for( idx = 0; idx < num_exps; idx++ ) { 894 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 895 } 896 897 // Add in mappings of the monitors 898 assert( !method || 899 !method->is_synchronized() || 900 method->is_native() || 901 num_mon > 0 || 902 !GenerateSynchronizationCode, 903 "monitors must always exist for synchronized methods"); 904 905 // Build the growable array of ScopeValues for exp stack 906 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 907 908 // Loop over monitors and insert into array 909 for (idx = 0; idx < num_mon; idx++) { 910 // Grab the node that defines this monitor 911 Node* box_node = sfn->monitor_box(jvms, idx); 912 Node* obj_node = sfn->monitor_obj(jvms, idx); 913 914 // Create ScopeValue for object 915 ScopeValue *scval = NULL; 916 917 if (obj_node->is_SafePointScalarObject()) { 918 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 919 scval = Compile::sv_for_node_id(objs, spobj->_idx); 920 if (scval == NULL) { 921 const Type *t = spobj->bottom_type(); 922 ciKlass* cik = t->is_oopptr()->klass(); 923 assert(cik->is_instance_klass() || 924 cik->is_array_klass(), "Not supported allocation."); 925 ObjectValue* sv = new ObjectValue(spobj->_idx, 926 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 927 Compile::set_sv_for_object_node(objs, sv); 928 929 uint first_ind = spobj->first_index(youngest_jvms); 930 for (uint i = 0; i < spobj->n_fields(); i++) { 931 Node* fld_node = sfn->in(first_ind+i); 932 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 933 } 934 scval = sv; 935 } 936 } else if (!obj_node->is_Con()) { 937 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 938 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 939 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 940 } else { 941 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 942 } 943 } else { 944 const TypePtr *tp = obj_node->get_ptr_type(); 945 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 946 } 947 948 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 949 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 950 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 951 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 952 } 953 954 // We dump the object pool first, since deoptimization reads it in first. 955 debug_info()->dump_object_pool(objs); 956 957 // Build first class objects to pass to scope 958 DebugToken *locvals = debug_info()->create_scope_values(locarray); 959 DebugToken *expvals = debug_info()->create_scope_values(exparray); 960 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 961 962 // Make method available for all Safepoints 963 ciMethod* scope_method = method ? method : _method; 964 // Describe the scope here 965 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 966 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 967 // Now we can describe the scope. 968 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals); 969 } // End jvms loop 970 971 // Mark the end of the scope set. 972 debug_info()->end_safepoint(safepoint_pc_offset); 973 } 974 975 976 977 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 978 class NonSafepointEmitter { 979 Compile* C; 980 JVMState* _pending_jvms; 981 int _pending_offset; 982 983 void emit_non_safepoint(); 984 985 public: 986 NonSafepointEmitter(Compile* compile) { 987 this->C = compile; 988 _pending_jvms = NULL; 989 _pending_offset = 0; 990 } 991 992 void observe_instruction(Node* n, int pc_offset) { 993 if (!C->debug_info()->recording_non_safepoints()) return; 994 995 Node_Notes* nn = C->node_notes_at(n->_idx); 996 if (nn == NULL || nn->jvms() == NULL) return; 997 if (_pending_jvms != NULL && 998 _pending_jvms->same_calls_as(nn->jvms())) { 999 // Repeated JVMS? Stretch it up here. 1000 _pending_offset = pc_offset; 1001 } else { 1002 if (_pending_jvms != NULL && 1003 _pending_offset < pc_offset) { 1004 emit_non_safepoint(); 1005 } 1006 _pending_jvms = NULL; 1007 if (pc_offset > C->debug_info()->last_pc_offset()) { 1008 // This is the only way _pending_jvms can become non-NULL: 1009 _pending_jvms = nn->jvms(); 1010 _pending_offset = pc_offset; 1011 } 1012 } 1013 } 1014 1015 // Stay out of the way of real safepoints: 1016 void observe_safepoint(JVMState* jvms, int pc_offset) { 1017 if (_pending_jvms != NULL && 1018 !_pending_jvms->same_calls_as(jvms) && 1019 _pending_offset < pc_offset) { 1020 emit_non_safepoint(); 1021 } 1022 _pending_jvms = NULL; 1023 } 1024 1025 void flush_at_end() { 1026 if (_pending_jvms != NULL) { 1027 emit_non_safepoint(); 1028 } 1029 _pending_jvms = NULL; 1030 } 1031 }; 1032 1033 void NonSafepointEmitter::emit_non_safepoint() { 1034 JVMState* youngest_jvms = _pending_jvms; 1035 int pc_offset = _pending_offset; 1036 1037 // Clear it now: 1038 _pending_jvms = NULL; 1039 1040 DebugInformationRecorder* debug_info = C->debug_info(); 1041 assert(debug_info->recording_non_safepoints(), "sanity"); 1042 1043 debug_info->add_non_safepoint(pc_offset); 1044 int max_depth = youngest_jvms->depth(); 1045 1046 // Visit scopes from oldest to youngest. 1047 for (int depth = 1; depth <= max_depth; depth++) { 1048 JVMState* jvms = youngest_jvms->of_depth(depth); 1049 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 1050 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1051 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute()); 1052 } 1053 1054 // Mark the end of the scope set. 1055 debug_info->end_non_safepoint(pc_offset); 1056 } 1057 1058 //------------------------------init_buffer------------------------------------ 1059 CodeBuffer* Compile::init_buffer(uint* blk_starts) { 1060 1061 // Set the initially allocated size 1062 int code_req = initial_code_capacity; 1063 int locs_req = initial_locs_capacity; 1064 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; 1065 int const_req = initial_const_capacity; 1066 1067 int pad_req = NativeCall::instruction_size; 1068 // The extra spacing after the code is necessary on some platforms. 1069 // Sometimes we need to patch in a jump after the last instruction, 1070 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1071 1072 // Compute the byte offset where we can store the deopt pc. 1073 if (fixed_slots() != 0) { 1074 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1075 } 1076 1077 // Compute prolog code size 1078 _method_size = 0; 1079 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 1080 #if defined(IA64) && !defined(AIX) 1081 if (save_argument_registers()) { 1082 // 4815101: this is a stub with implicit and unknown precision fp args. 1083 // The usual spill mechanism can only generate stfd's in this case, which 1084 // doesn't work if the fp reg to spill contains a single-precision denorm. 1085 // Instead, we hack around the normal spill mechanism using stfspill's and 1086 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 1087 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 1088 // 1089 // If we ever implement 16-byte 'registers' == stack slots, we can 1090 // get rid of this hack and have SpillCopy generate stfspill/ldffill 1091 // instead of stfd/stfs/ldfd/ldfs. 1092 _frame_slots += 8*(16/BytesPerInt); 1093 } 1094 #endif 1095 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 1096 1097 if (has_mach_constant_base_node()) { 1098 uint add_size = 0; 1099 // Fill the constant table. 1100 // Note: This must happen before shorten_branches. 1101 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1102 Block* b = _cfg->get_block(i); 1103 1104 for (uint j = 0; j < b->number_of_nodes(); j++) { 1105 Node* n = b->get_node(j); 1106 1107 // If the node is a MachConstantNode evaluate the constant 1108 // value section. 1109 if (n->is_MachConstant()) { 1110 MachConstantNode* machcon = n->as_MachConstant(); 1111 machcon->eval_constant(C); 1112 } else if (n->is_Mach()) { 1113 // On Power there are more nodes that issue constants. 1114 add_size += (n->as_Mach()->ins_num_consts() * 8); 1115 } 1116 } 1117 } 1118 1119 // Calculate the offsets of the constants and the size of the 1120 // constant table (including the padding to the next section). 1121 constant_table().calculate_offsets_and_size(); 1122 const_req = constant_table().size() + add_size; 1123 } 1124 1125 // Initialize the space for the BufferBlob used to find and verify 1126 // instruction size in MachNode::emit_size() 1127 init_scratch_buffer_blob(const_req); 1128 if (failing()) return NULL; // Out of memory 1129 1130 // Pre-compute the length of blocks and replace 1131 // long branches with short if machine supports it. 1132 shorten_branches(blk_starts, code_req, locs_req, stub_req); 1133 1134 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1135 // class HandlerImpl is platform-specific and defined in the *.ad files. 1136 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler 1137 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler 1138 stub_req += MAX_stubs_size; // ensure per-stub margin 1139 code_req += MAX_inst_size; // ensure per-instruction margin 1140 1141 if (StressCodeBuffers) 1142 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1143 1144 int total_req = 1145 const_req + 1146 code_req + 1147 pad_req + 1148 stub_req + 1149 exception_handler_req + 1150 deopt_handler_req; // deopt handler 1151 1152 if (has_method_handle_invokes()) 1153 total_req += deopt_handler_req; // deopt MH handler 1154 1155 CodeBuffer* cb = code_buffer(); 1156 cb->initialize(total_req, locs_req); 1157 1158 // Have we run out of code space? 1159 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1160 C->record_failure("CodeCache is full"); 1161 return NULL; 1162 } 1163 // Configure the code buffer. 1164 cb->initialize_consts_size(const_req); 1165 cb->initialize_stubs_size(stub_req); 1166 cb->initialize_oop_recorder(env()->oop_recorder()); 1167 1168 // fill in the nop array for bundling computations 1169 MachNode *_nop_list[Bundle::_nop_count]; 1170 Bundle::initialize_nops(_nop_list, this); 1171 1172 return cb; 1173 } 1174 1175 //------------------------------fill_buffer------------------------------------ 1176 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1177 // blk_starts[] contains offsets calculated during short branches processing, 1178 // offsets should not be increased during following steps. 1179 1180 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1181 // of a loop. It is used to determine the padding for loop alignment. 1182 compute_loop_first_inst_sizes(); 1183 1184 // Create oopmap set. 1185 _oop_map_set = new OopMapSet(); 1186 1187 // !!!!! This preserves old handling of oopmaps for now 1188 debug_info()->set_oopmaps(_oop_map_set); 1189 1190 uint nblocks = _cfg->number_of_blocks(); 1191 // Count and start of implicit null check instructions 1192 uint inct_cnt = 0; 1193 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1194 1195 // Count and start of calls 1196 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1197 1198 uint return_offset = 0; 1199 int nop_size = (new (this) MachNopNode())->size(_regalloc); 1200 1201 int previous_offset = 0; 1202 int current_offset = 0; 1203 int last_call_offset = -1; 1204 int last_avoid_back_to_back_offset = -1; 1205 #ifdef ASSERT 1206 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1207 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1208 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1209 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1210 #endif 1211 1212 // Create an array of unused labels, one for each basic block, if printing is enabled 1213 #ifndef PRODUCT 1214 int *node_offsets = NULL; 1215 uint node_offset_limit = unique(); 1216 1217 if (print_assembly()) 1218 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1219 #endif 1220 1221 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1222 1223 // Emit the constant table. 1224 if (has_mach_constant_base_node()) { 1225 constant_table().emit(*cb); 1226 } 1227 1228 // Create an array of labels, one for each basic block 1229 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1230 for (uint i=0; i <= nblocks; i++) { 1231 blk_labels[i].init(); 1232 } 1233 1234 // ------------------ 1235 // Now fill in the code buffer 1236 Node *delay_slot = NULL; 1237 1238 for (uint i = 0; i < nblocks; i++) { 1239 Block* block = _cfg->get_block(i); 1240 Node* head = block->head(); 1241 1242 // If this block needs to start aligned (i.e, can be reached other 1243 // than by falling-thru from the previous block), then force the 1244 // start of a new bundle. 1245 if (Pipeline::requires_bundling() && starts_bundle(head)) { 1246 cb->flush_bundle(true); 1247 } 1248 1249 #ifdef ASSERT 1250 if (!block->is_connector()) { 1251 stringStream st; 1252 block->dump_head(_cfg, &st); 1253 MacroAssembler(cb).block_comment(st.as_string()); 1254 } 1255 jmp_target[i] = 0; 1256 jmp_offset[i] = 0; 1257 jmp_size[i] = 0; 1258 jmp_rule[i] = 0; 1259 #endif 1260 int blk_offset = current_offset; 1261 1262 // Define the label at the beginning of the basic block 1263 MacroAssembler(cb).bind(blk_labels[block->_pre_order]); 1264 1265 uint last_inst = block->number_of_nodes(); 1266 1267 // Emit block normally, except for last instruction. 1268 // Emit means "dump code bits into code buffer". 1269 for (uint j = 0; j<last_inst; j++) { 1270 1271 // Get the node 1272 Node* n = block->get_node(j); 1273 1274 // See if delay slots are supported 1275 if (valid_bundle_info(n) && 1276 node_bundling(n)->used_in_unconditional_delay()) { 1277 assert(delay_slot == NULL, "no use of delay slot node"); 1278 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1279 1280 delay_slot = n; 1281 continue; 1282 } 1283 1284 // If this starts a new instruction group, then flush the current one 1285 // (but allow split bundles) 1286 if (Pipeline::requires_bundling() && starts_bundle(n)) 1287 cb->flush_bundle(false); 1288 1289 // The following logic is duplicated in the code ifdeffed for 1290 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It 1291 // should be factored out. Or maybe dispersed to the nodes? 1292 1293 // Special handling for SafePoint/Call Nodes 1294 bool is_mcall = false; 1295 if (n->is_Mach()) { 1296 MachNode *mach = n->as_Mach(); 1297 is_mcall = n->is_MachCall(); 1298 bool is_sfn = n->is_MachSafePoint(); 1299 1300 // If this requires all previous instructions be flushed, then do so 1301 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1302 cb->flush_bundle(true); 1303 current_offset = cb->insts_size(); 1304 } 1305 1306 // A padding may be needed again since a previous instruction 1307 // could be moved to delay slot. 1308 1309 // align the instruction if necessary 1310 int padding = mach->compute_padding(current_offset); 1311 // Make sure safepoint node for polling is distinct from a call's 1312 // return by adding a nop if needed. 1313 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1314 padding = nop_size; 1315 } 1316 if (padding == 0 && mach->avoid_back_to_back() && 1317 current_offset == last_avoid_back_to_back_offset) { 1318 // Avoid back to back some instructions. 1319 padding = nop_size; 1320 } 1321 1322 if(padding > 0) { 1323 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1324 int nops_cnt = padding / nop_size; 1325 MachNode *nop = new (this) MachNopNode(nops_cnt); 1326 block->insert_node(nop, j++); 1327 last_inst++; 1328 _cfg->map_node_to_block(nop, block); 1329 nop->emit(*cb, _regalloc); 1330 cb->flush_bundle(true); 1331 current_offset = cb->insts_size(); 1332 } 1333 1334 // Remember the start of the last call in a basic block 1335 if (is_mcall) { 1336 MachCallNode *mcall = mach->as_MachCall(); 1337 1338 // This destination address is NOT PC-relative 1339 mcall->method_set((intptr_t)mcall->entry_point()); 1340 1341 // Save the return address 1342 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset(); 1343 1344 if (mcall->is_MachCallLeaf()) { 1345 is_mcall = false; 1346 is_sfn = false; 1347 } 1348 } 1349 1350 // sfn will be valid whenever mcall is valid now because of inheritance 1351 if (is_sfn || is_mcall) { 1352 1353 // Handle special safepoint nodes for synchronization 1354 if (!is_mcall) { 1355 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1356 // !!!!! Stubs only need an oopmap right now, so bail out 1357 if (sfn->jvms()->method() == NULL) { 1358 // Write the oopmap directly to the code blob??!! 1359 # ifdef ENABLE_ZAP_DEAD_LOCALS 1360 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive"); 1361 # endif 1362 continue; 1363 } 1364 } // End synchronization 1365 1366 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1367 current_offset); 1368 Process_OopMap_Node(mach, current_offset); 1369 } // End if safepoint 1370 1371 // If this is a null check, then add the start of the previous instruction to the list 1372 else if( mach->is_MachNullCheck() ) { 1373 inct_starts[inct_cnt++] = previous_offset; 1374 } 1375 1376 // If this is a branch, then fill in the label with the target BB's label 1377 else if (mach->is_MachBranch()) { 1378 // This requires the TRUE branch target be in succs[0] 1379 uint block_num = block->non_connector_successor(0)->_pre_order; 1380 1381 // Try to replace long branch if delay slot is not used, 1382 // it is mostly for back branches since forward branch's 1383 // distance is not updated yet. 1384 bool delay_slot_is_used = valid_bundle_info(n) && 1385 node_bundling(n)->use_unconditional_delay(); 1386 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1387 assert(delay_slot == NULL, "not expecting delay slot node"); 1388 int br_size = n->size(_regalloc); 1389 int offset = blk_starts[block_num] - current_offset; 1390 if (block_num >= i) { 1391 // Current and following block's offset are not 1392 // finalized yet, adjust distance by the difference 1393 // between calculated and final offsets of current block. 1394 offset -= (blk_starts[i] - blk_offset); 1395 } 1396 // In the following code a nop could be inserted before 1397 // the branch which will increase the backward distance. 1398 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1399 if (needs_padding && offset <= 0) 1400 offset -= nop_size; 1401 1402 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 1403 // We've got a winner. Replace this branch. 1404 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 1405 1406 // Update the jmp_size. 1407 int new_size = replacement->size(_regalloc); 1408 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1409 // Insert padding between avoid_back_to_back branches. 1410 if (needs_padding && replacement->avoid_back_to_back()) { 1411 MachNode *nop = new (this) MachNopNode(); 1412 block->insert_node(nop, j++); 1413 _cfg->map_node_to_block(nop, block); 1414 last_inst++; 1415 nop->emit(*cb, _regalloc); 1416 cb->flush_bundle(true); 1417 current_offset = cb->insts_size(); 1418 } 1419 #ifdef ASSERT 1420 jmp_target[i] = block_num; 1421 jmp_offset[i] = current_offset - blk_offset; 1422 jmp_size[i] = new_size; 1423 jmp_rule[i] = mach->rule(); 1424 #endif 1425 block->map_node(replacement, j); 1426 mach->subsume_by(replacement, C); 1427 n = replacement; 1428 mach = replacement; 1429 } 1430 } 1431 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1432 } else if (mach->ideal_Opcode() == Op_Jump) { 1433 for (uint h = 0; h < block->_num_succs; h++) { 1434 Block* succs_block = block->_succs[h]; 1435 for (uint j = 1; j < succs_block->num_preds(); j++) { 1436 Node* jpn = succs_block->pred(j); 1437 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1438 uint block_num = succs_block->non_connector()->_pre_order; 1439 Label *blkLabel = &blk_labels[block_num]; 1440 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1441 } 1442 } 1443 } 1444 } 1445 #ifdef ASSERT 1446 // Check that oop-store precedes the card-mark 1447 else if (mach->ideal_Opcode() == Op_StoreCM) { 1448 uint storeCM_idx = j; 1449 int count = 0; 1450 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1451 Node *oop_store = mach->in(prec); // Precedence edge 1452 if (oop_store == NULL) continue; 1453 count++; 1454 uint i4; 1455 for (i4 = 0; i4 < last_inst; ++i4) { 1456 if (block->get_node(i4) == oop_store) { 1457 break; 1458 } 1459 } 1460 // Note: This test can provide a false failure if other precedence 1461 // edges have been added to the storeCMNode. 1462 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1463 } 1464 assert(count > 0, "storeCM expects at least one precedence edge"); 1465 } 1466 #endif 1467 else if (!n->is_Proj()) { 1468 // Remember the beginning of the previous instruction, in case 1469 // it's followed by a flag-kill and a null-check. Happens on 1470 // Intel all the time, with add-to-memory kind of opcodes. 1471 previous_offset = current_offset; 1472 } 1473 1474 // Not an else-if! 1475 // If this is a trap based cmp then add its offset to the list. 1476 if (mach->is_TrapBasedCheckNode()) { 1477 inct_starts[inct_cnt++] = current_offset; 1478 } 1479 } 1480 1481 // Verify that there is sufficient space remaining 1482 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1483 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1484 C->record_failure("CodeCache is full"); 1485 return; 1486 } 1487 1488 // Save the offset for the listing 1489 #ifndef PRODUCT 1490 if (node_offsets && n->_idx < node_offset_limit) 1491 node_offsets[n->_idx] = cb->insts_size(); 1492 #endif 1493 1494 // "Normal" instruction case 1495 DEBUG_ONLY( uint instr_offset = cb->insts_size(); ) 1496 n->emit(*cb, _regalloc); 1497 current_offset = cb->insts_size(); 1498 1499 #ifdef ASSERT 1500 if (n->size(_regalloc) < (current_offset-instr_offset)) { 1501 n->dump(); 1502 assert(false, "wrong size of mach node"); 1503 } 1504 #endif 1505 non_safepoints.observe_instruction(n, current_offset); 1506 1507 // mcall is last "call" that can be a safepoint 1508 // record it so we can see if a poll will directly follow it 1509 // in which case we'll need a pad to make the PcDesc sites unique 1510 // see 5010568. This can be slightly inaccurate but conservative 1511 // in the case that return address is not actually at current_offset. 1512 // This is a small price to pay. 1513 1514 if (is_mcall) { 1515 last_call_offset = current_offset; 1516 } 1517 1518 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) { 1519 // Avoid back to back some instructions. 1520 last_avoid_back_to_back_offset = current_offset; 1521 } 1522 1523 // See if this instruction has a delay slot 1524 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1525 assert(delay_slot != NULL, "expecting delay slot node"); 1526 1527 // Back up 1 instruction 1528 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1529 1530 // Save the offset for the listing 1531 #ifndef PRODUCT 1532 if (node_offsets && delay_slot->_idx < node_offset_limit) 1533 node_offsets[delay_slot->_idx] = cb->insts_size(); 1534 #endif 1535 1536 // Support a SafePoint in the delay slot 1537 if (delay_slot->is_MachSafePoint()) { 1538 MachNode *mach = delay_slot->as_Mach(); 1539 // !!!!! Stubs only need an oopmap right now, so bail out 1540 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) { 1541 // Write the oopmap directly to the code blob??!! 1542 # ifdef ENABLE_ZAP_DEAD_LOCALS 1543 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive"); 1544 # endif 1545 delay_slot = NULL; 1546 continue; 1547 } 1548 1549 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1550 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1551 adjusted_offset); 1552 // Generate an OopMap entry 1553 Process_OopMap_Node(mach, adjusted_offset); 1554 } 1555 1556 // Insert the delay slot instruction 1557 delay_slot->emit(*cb, _regalloc); 1558 1559 // Don't reuse it 1560 delay_slot = NULL; 1561 } 1562 1563 } // End for all instructions in block 1564 1565 // If the next block is the top of a loop, pad this block out to align 1566 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1567 if (i < nblocks-1) { 1568 Block *nb = _cfg->get_block(i + 1); 1569 int padding = nb->alignment_padding(current_offset); 1570 if( padding > 0 ) { 1571 MachNode *nop = new (this) MachNopNode(padding / nop_size); 1572 block->insert_node(nop, block->number_of_nodes()); 1573 _cfg->map_node_to_block(nop, block); 1574 nop->emit(*cb, _regalloc); 1575 current_offset = cb->insts_size(); 1576 } 1577 } 1578 // Verify that the distance for generated before forward 1579 // short branches is still valid. 1580 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1581 1582 // Save new block start offset 1583 blk_starts[i] = blk_offset; 1584 } // End of for all blocks 1585 blk_starts[nblocks] = current_offset; 1586 1587 non_safepoints.flush_at_end(); 1588 1589 // Offset too large? 1590 if (failing()) return; 1591 1592 // Define a pseudo-label at the end of the code 1593 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1594 1595 // Compute the size of the first block 1596 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1597 1598 assert(cb->insts_size() < 500000, "method is unreasonably large"); 1599 1600 #ifdef ASSERT 1601 for (uint i = 0; i < nblocks; i++) { // For all blocks 1602 if (jmp_target[i] != 0) { 1603 int br_size = jmp_size[i]; 1604 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1605 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1606 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]); 1607 assert(false, "Displacement too large for short jmp"); 1608 } 1609 } 1610 } 1611 #endif 1612 1613 #ifndef PRODUCT 1614 // Information on the size of the method, without the extraneous code 1615 Scheduling::increment_method_size(cb->insts_size()); 1616 #endif 1617 1618 // ------------------ 1619 // Fill in exception table entries. 1620 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1621 1622 // Only java methods have exception handlers and deopt handlers 1623 // class HandlerImpl is platform-specific and defined in the *.ad files. 1624 if (_method) { 1625 // Emit the exception handler code. 1626 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb)); 1627 // Emit the deopt handler code. 1628 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb)); 1629 1630 // Emit the MethodHandle deopt handler code (if required). 1631 if (has_method_handle_invokes()) { 1632 // We can use the same code as for the normal deopt handler, we 1633 // just need a different entry point address. 1634 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb)); 1635 } 1636 } 1637 1638 // One last check for failed CodeBuffer::expand: 1639 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1640 C->record_failure("CodeCache is full"); 1641 return; 1642 } 1643 1644 #ifndef PRODUCT 1645 // Dump the assembly code, including basic-block numbers 1646 if (print_assembly()) { 1647 ttyLocker ttyl; // keep the following output all in one block 1648 if (!VMThread::should_terminate()) { // test this under the tty lock 1649 // This output goes directly to the tty, not the compiler log. 1650 // To enable tools to match it up with the compilation activity, 1651 // be sure to tag this tty output with the compile ID. 1652 if (xtty != NULL) { 1653 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1654 is_osr_compilation() ? " compile_kind='osr'" : 1655 ""); 1656 } 1657 if (method() != NULL) { 1658 method()->print_metadata(); 1659 } 1660 dump_asm(node_offsets, node_offset_limit); 1661 if (xtty != NULL) { 1662 xtty->tail("opto_assembly"); 1663 } 1664 } 1665 } 1666 #endif 1667 1668 } 1669 1670 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1671 _inc_table.set_size(cnt); 1672 1673 uint inct_cnt = 0; 1674 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1675 Block* block = _cfg->get_block(i); 1676 Node *n = NULL; 1677 int j; 1678 1679 // Find the branch; ignore trailing NOPs. 1680 for (j = block->number_of_nodes() - 1; j >= 0; j--) { 1681 n = block->get_node(j); 1682 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) { 1683 break; 1684 } 1685 } 1686 1687 // If we didn't find anything, continue 1688 if (j < 0) { 1689 continue; 1690 } 1691 1692 // Compute ExceptionHandlerTable subtable entry and add it 1693 // (skip empty blocks) 1694 if (n->is_Catch()) { 1695 1696 // Get the offset of the return from the call 1697 uint call_return = call_returns[block->_pre_order]; 1698 #ifdef ASSERT 1699 assert( call_return > 0, "no call seen for this basic block" ); 1700 while (block->get_node(--j)->is_MachProj()) ; 1701 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1702 #endif 1703 // last instruction is a CatchNode, find it's CatchProjNodes 1704 int nof_succs = block->_num_succs; 1705 // allocate space 1706 GrowableArray<intptr_t> handler_bcis(nof_succs); 1707 GrowableArray<intptr_t> handler_pcos(nof_succs); 1708 // iterate through all successors 1709 for (int j = 0; j < nof_succs; j++) { 1710 Block* s = block->_succs[j]; 1711 bool found_p = false; 1712 for (uint k = 1; k < s->num_preds(); k++) { 1713 Node* pk = s->pred(k); 1714 if (pk->is_CatchProj() && pk->in(0) == n) { 1715 const CatchProjNode* p = pk->as_CatchProj(); 1716 found_p = true; 1717 // add the corresponding handler bci & pco information 1718 if (p->_con != CatchProjNode::fall_through_index) { 1719 // p leads to an exception handler (and is not fall through) 1720 assert(s == _cfg->get_block(s->_pre_order), "bad numbering"); 1721 // no duplicates, please 1722 if (!handler_bcis.contains(p->handler_bci())) { 1723 uint block_num = s->non_connector()->_pre_order; 1724 handler_bcis.append(p->handler_bci()); 1725 handler_pcos.append(blk_labels[block_num].loc_pos()); 1726 } 1727 } 1728 } 1729 } 1730 assert(found_p, "no matching predecessor found"); 1731 // Note: Due to empty block removal, one block may have 1732 // several CatchProj inputs, from the same Catch. 1733 } 1734 1735 // Set the offset of the return from the call 1736 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1737 continue; 1738 } 1739 1740 // Handle implicit null exception table updates 1741 if (n->is_MachNullCheck()) { 1742 uint block_num = block->non_connector_successor(0)->_pre_order; 1743 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1744 continue; 1745 } 1746 // Handle implicit exception table updates: trap instructions. 1747 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) { 1748 uint block_num = block->non_connector_successor(0)->_pre_order; 1749 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1750 continue; 1751 } 1752 } // End of for all blocks fill in exception table entries 1753 } 1754 1755 // Static Variables 1756 #ifndef PRODUCT 1757 uint Scheduling::_total_nop_size = 0; 1758 uint Scheduling::_total_method_size = 0; 1759 uint Scheduling::_total_branches = 0; 1760 uint Scheduling::_total_unconditional_delays = 0; 1761 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1762 #endif 1763 1764 // Initializer for class Scheduling 1765 1766 Scheduling::Scheduling(Arena *arena, Compile &compile) 1767 : _arena(arena), 1768 _cfg(compile.cfg()), 1769 _regalloc(compile.regalloc()), 1770 _reg_node(arena), 1771 _bundle_instr_count(0), 1772 _bundle_cycle_number(0), 1773 _scheduled(arena), 1774 _available(arena), 1775 _next_node(NULL), 1776 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1777 _pinch_free_list(arena) 1778 #ifndef PRODUCT 1779 , _branches(0) 1780 , _unconditional_delays(0) 1781 #endif 1782 { 1783 // Create a MachNopNode 1784 _nop = new (&compile) MachNopNode(); 1785 1786 // Now that the nops are in the array, save the count 1787 // (but allow entries for the nops) 1788 _node_bundling_limit = compile.unique(); 1789 uint node_max = _regalloc->node_regs_max_index(); 1790 1791 compile.set_node_bundling_limit(_node_bundling_limit); 1792 1793 // This one is persistent within the Compile class 1794 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1795 1796 // Allocate space for fixed-size arrays 1797 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1798 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1799 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1800 1801 // Clear the arrays 1802 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1803 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1804 memset(_uses, 0, node_max * sizeof(short)); 1805 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1806 1807 // Clear the bundling information 1808 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements)); 1809 1810 // Get the last node 1811 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1); 1812 1813 _next_node = block->get_node(block->number_of_nodes() - 1); 1814 } 1815 1816 #ifndef PRODUCT 1817 // Scheduling destructor 1818 Scheduling::~Scheduling() { 1819 _total_branches += _branches; 1820 _total_unconditional_delays += _unconditional_delays; 1821 } 1822 #endif 1823 1824 // Step ahead "i" cycles 1825 void Scheduling::step(uint i) { 1826 1827 Bundle *bundle = node_bundling(_next_node); 1828 bundle->set_starts_bundle(); 1829 1830 // Update the bundle record, but leave the flags information alone 1831 if (_bundle_instr_count > 0) { 1832 bundle->set_instr_count(_bundle_instr_count); 1833 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1834 } 1835 1836 // Update the state information 1837 _bundle_instr_count = 0; 1838 _bundle_cycle_number += i; 1839 _bundle_use.step(i); 1840 } 1841 1842 void Scheduling::step_and_clear() { 1843 Bundle *bundle = node_bundling(_next_node); 1844 bundle->set_starts_bundle(); 1845 1846 // Update the bundle record 1847 if (_bundle_instr_count > 0) { 1848 bundle->set_instr_count(_bundle_instr_count); 1849 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1850 1851 _bundle_cycle_number += 1; 1852 } 1853 1854 // Clear the bundling information 1855 _bundle_instr_count = 0; 1856 _bundle_use.reset(); 1857 1858 memcpy(_bundle_use_elements, 1859 Pipeline_Use::elaborated_elements, 1860 sizeof(Pipeline_Use::elaborated_elements)); 1861 } 1862 1863 // Perform instruction scheduling and bundling over the sequence of 1864 // instructions in backwards order. 1865 void Compile::ScheduleAndBundle() { 1866 1867 // Don't optimize this if it isn't a method 1868 if (!_method) 1869 return; 1870 1871 // Don't optimize this if scheduling is disabled 1872 if (!do_scheduling()) 1873 return; 1874 1875 // Scheduling code works only with pairs (8 bytes) maximum. 1876 if (max_vector_size() > 8) 1877 return; 1878 1879 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); ) 1880 1881 // Create a data structure for all the scheduling information 1882 Scheduling scheduling(Thread::current()->resource_area(), *this); 1883 1884 // Walk backwards over each basic block, computing the needed alignment 1885 // Walk over all the basic blocks 1886 scheduling.DoScheduling(); 1887 } 1888 1889 // Compute the latency of all the instructions. This is fairly simple, 1890 // because we already have a legal ordering. Walk over the instructions 1891 // from first to last, and compute the latency of the instruction based 1892 // on the latency of the preceding instruction(s). 1893 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1894 #ifndef PRODUCT 1895 if (_cfg->C->trace_opto_output()) 1896 tty->print("# -> ComputeLocalLatenciesForward\n"); 1897 #endif 1898 1899 // Walk over all the schedulable instructions 1900 for( uint j=_bb_start; j < _bb_end; j++ ) { 1901 1902 // This is a kludge, forcing all latency calculations to start at 1. 1903 // Used to allow latency 0 to force an instruction to the beginning 1904 // of the bb 1905 uint latency = 1; 1906 Node *use = bb->get_node(j); 1907 uint nlen = use->len(); 1908 1909 // Walk over all the inputs 1910 for ( uint k=0; k < nlen; k++ ) { 1911 Node *def = use->in(k); 1912 if (!def) 1913 continue; 1914 1915 uint l = _node_latency[def->_idx] + use->latency(k); 1916 if (latency < l) 1917 latency = l; 1918 } 1919 1920 _node_latency[use->_idx] = latency; 1921 1922 #ifndef PRODUCT 1923 if (_cfg->C->trace_opto_output()) { 1924 tty->print("# latency %4d: ", latency); 1925 use->dump(); 1926 } 1927 #endif 1928 } 1929 1930 #ifndef PRODUCT 1931 if (_cfg->C->trace_opto_output()) 1932 tty->print("# <- ComputeLocalLatenciesForward\n"); 1933 #endif 1934 1935 } // end ComputeLocalLatenciesForward 1936 1937 // See if this node fits into the present instruction bundle 1938 bool Scheduling::NodeFitsInBundle(Node *n) { 1939 uint n_idx = n->_idx; 1940 1941 // If this is the unconditional delay instruction, then it fits 1942 if (n == _unconditional_delay_slot) { 1943 #ifndef PRODUCT 1944 if (_cfg->C->trace_opto_output()) 1945 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1946 #endif 1947 return (true); 1948 } 1949 1950 // If the node cannot be scheduled this cycle, skip it 1951 if (_current_latency[n_idx] > _bundle_cycle_number) { 1952 #ifndef PRODUCT 1953 if (_cfg->C->trace_opto_output()) 1954 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1955 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1956 #endif 1957 return (false); 1958 } 1959 1960 const Pipeline *node_pipeline = n->pipeline(); 1961 1962 uint instruction_count = node_pipeline->instructionCount(); 1963 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1964 instruction_count = 0; 1965 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1966 instruction_count++; 1967 1968 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1969 #ifndef PRODUCT 1970 if (_cfg->C->trace_opto_output()) 1971 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1972 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1973 #endif 1974 return (false); 1975 } 1976 1977 // Don't allow non-machine nodes to be handled this way 1978 if (!n->is_Mach() && instruction_count == 0) 1979 return (false); 1980 1981 // See if there is any overlap 1982 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1983 1984 if (delay > 0) { 1985 #ifndef PRODUCT 1986 if (_cfg->C->trace_opto_output()) 1987 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1988 #endif 1989 return false; 1990 } 1991 1992 #ifndef PRODUCT 1993 if (_cfg->C->trace_opto_output()) 1994 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1995 #endif 1996 1997 return true; 1998 } 1999 2000 Node * Scheduling::ChooseNodeToBundle() { 2001 uint siz = _available.size(); 2002 2003 if (siz == 0) { 2004 2005 #ifndef PRODUCT 2006 if (_cfg->C->trace_opto_output()) 2007 tty->print("# ChooseNodeToBundle: NULL\n"); 2008 #endif 2009 return (NULL); 2010 } 2011 2012 // Fast path, if only 1 instruction in the bundle 2013 if (siz == 1) { 2014 #ifndef PRODUCT 2015 if (_cfg->C->trace_opto_output()) { 2016 tty->print("# ChooseNodeToBundle (only 1): "); 2017 _available[0]->dump(); 2018 } 2019 #endif 2020 return (_available[0]); 2021 } 2022 2023 // Don't bother, if the bundle is already full 2024 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 2025 for ( uint i = 0; i < siz; i++ ) { 2026 Node *n = _available[i]; 2027 2028 // Skip projections, we'll handle them another way 2029 if (n->is_Proj()) 2030 continue; 2031 2032 // This presupposed that instructions are inserted into the 2033 // available list in a legality order; i.e. instructions that 2034 // must be inserted first are at the head of the list 2035 if (NodeFitsInBundle(n)) { 2036 #ifndef PRODUCT 2037 if (_cfg->C->trace_opto_output()) { 2038 tty->print("# ChooseNodeToBundle: "); 2039 n->dump(); 2040 } 2041 #endif 2042 return (n); 2043 } 2044 } 2045 } 2046 2047 // Nothing fits in this bundle, choose the highest priority 2048 #ifndef PRODUCT 2049 if (_cfg->C->trace_opto_output()) { 2050 tty->print("# ChooseNodeToBundle: "); 2051 _available[0]->dump(); 2052 } 2053 #endif 2054 2055 return _available[0]; 2056 } 2057 2058 void Scheduling::AddNodeToAvailableList(Node *n) { 2059 assert( !n->is_Proj(), "projections never directly made available" ); 2060 #ifndef PRODUCT 2061 if (_cfg->C->trace_opto_output()) { 2062 tty->print("# AddNodeToAvailableList: "); 2063 n->dump(); 2064 } 2065 #endif 2066 2067 int latency = _current_latency[n->_idx]; 2068 2069 // Insert in latency order (insertion sort) 2070 uint i; 2071 for ( i=0; i < _available.size(); i++ ) 2072 if (_current_latency[_available[i]->_idx] > latency) 2073 break; 2074 2075 // Special Check for compares following branches 2076 if( n->is_Mach() && _scheduled.size() > 0 ) { 2077 int op = n->as_Mach()->ideal_Opcode(); 2078 Node *last = _scheduled[0]; 2079 if( last->is_MachIf() && last->in(1) == n && 2080 ( op == Op_CmpI || 2081 op == Op_CmpU || 2082 op == Op_CmpP || 2083 op == Op_CmpF || 2084 op == Op_CmpD || 2085 op == Op_CmpL ) ) { 2086 2087 // Recalculate position, moving to front of same latency 2088 for ( i=0 ; i < _available.size(); i++ ) 2089 if (_current_latency[_available[i]->_idx] >= latency) 2090 break; 2091 } 2092 } 2093 2094 // Insert the node in the available list 2095 _available.insert(i, n); 2096 2097 #ifndef PRODUCT 2098 if (_cfg->C->trace_opto_output()) 2099 dump_available(); 2100 #endif 2101 } 2102 2103 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2104 for ( uint i=0; i < n->len(); i++ ) { 2105 Node *def = n->in(i); 2106 if (!def) continue; 2107 if( def->is_Proj() ) // If this is a machine projection, then 2108 def = def->in(0); // propagate usage thru to the base instruction 2109 2110 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local 2111 continue; 2112 } 2113 2114 // Compute the latency 2115 uint l = _bundle_cycle_number + n->latency(i); 2116 if (_current_latency[def->_idx] < l) 2117 _current_latency[def->_idx] = l; 2118 2119 // If this does not have uses then schedule it 2120 if ((--_uses[def->_idx]) == 0) 2121 AddNodeToAvailableList(def); 2122 } 2123 } 2124 2125 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2126 #ifndef PRODUCT 2127 if (_cfg->C->trace_opto_output()) { 2128 tty->print("# AddNodeToBundle: "); 2129 n->dump(); 2130 } 2131 #endif 2132 2133 // Remove this from the available list 2134 uint i; 2135 for (i = 0; i < _available.size(); i++) 2136 if (_available[i] == n) 2137 break; 2138 assert(i < _available.size(), "entry in _available list not found"); 2139 _available.remove(i); 2140 2141 // See if this fits in the current bundle 2142 const Pipeline *node_pipeline = n->pipeline(); 2143 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2144 2145 // Check for instructions to be placed in the delay slot. We 2146 // do this before we actually schedule the current instruction, 2147 // because the delay slot follows the current instruction. 2148 if (Pipeline::_branch_has_delay_slot && 2149 node_pipeline->hasBranchDelay() && 2150 !_unconditional_delay_slot) { 2151 2152 uint siz = _available.size(); 2153 2154 // Conditional branches can support an instruction that 2155 // is unconditionally executed and not dependent by the 2156 // branch, OR a conditionally executed instruction if 2157 // the branch is taken. In practice, this means that 2158 // the first instruction at the branch target is 2159 // copied to the delay slot, and the branch goes to 2160 // the instruction after that at the branch target 2161 if ( n->is_MachBranch() ) { 2162 2163 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2164 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2165 2166 #ifndef PRODUCT 2167 _branches++; 2168 #endif 2169 2170 // At least 1 instruction is on the available list 2171 // that is not dependent on the branch 2172 for (uint i = 0; i < siz; i++) { 2173 Node *d = _available[i]; 2174 const Pipeline *avail_pipeline = d->pipeline(); 2175 2176 // Don't allow safepoints in the branch shadow, that will 2177 // cause a number of difficulties 2178 if ( avail_pipeline->instructionCount() == 1 && 2179 !avail_pipeline->hasMultipleBundles() && 2180 !avail_pipeline->hasBranchDelay() && 2181 Pipeline::instr_has_unit_size() && 2182 d->size(_regalloc) == Pipeline::instr_unit_size() && 2183 NodeFitsInBundle(d) && 2184 !node_bundling(d)->used_in_delay()) { 2185 2186 if (d->is_Mach() && !d->is_MachSafePoint()) { 2187 // A node that fits in the delay slot was found, so we need to 2188 // set the appropriate bits in the bundle pipeline information so 2189 // that it correctly indicates resource usage. Later, when we 2190 // attempt to add this instruction to the bundle, we will skip 2191 // setting the resource usage. 2192 _unconditional_delay_slot = d; 2193 node_bundling(n)->set_use_unconditional_delay(); 2194 node_bundling(d)->set_used_in_unconditional_delay(); 2195 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2196 _current_latency[d->_idx] = _bundle_cycle_number; 2197 _next_node = d; 2198 ++_bundle_instr_count; 2199 #ifndef PRODUCT 2200 _unconditional_delays++; 2201 #endif 2202 break; 2203 } 2204 } 2205 } 2206 } 2207 2208 // No delay slot, add a nop to the usage 2209 if (!_unconditional_delay_slot) { 2210 // See if adding an instruction in the delay slot will overflow 2211 // the bundle. 2212 if (!NodeFitsInBundle(_nop)) { 2213 #ifndef PRODUCT 2214 if (_cfg->C->trace_opto_output()) 2215 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2216 #endif 2217 step(1); 2218 } 2219 2220 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2221 _next_node = _nop; 2222 ++_bundle_instr_count; 2223 } 2224 2225 // See if the instruction in the delay slot requires a 2226 // step of the bundles 2227 if (!NodeFitsInBundle(n)) { 2228 #ifndef PRODUCT 2229 if (_cfg->C->trace_opto_output()) 2230 tty->print("# *** STEP(branch won't fit) ***\n"); 2231 #endif 2232 // Update the state information 2233 _bundle_instr_count = 0; 2234 _bundle_cycle_number += 1; 2235 _bundle_use.step(1); 2236 } 2237 } 2238 2239 // Get the number of instructions 2240 uint instruction_count = node_pipeline->instructionCount(); 2241 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2242 instruction_count = 0; 2243 2244 // Compute the latency information 2245 uint delay = 0; 2246 2247 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2248 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2249 if (relative_latency < 0) 2250 relative_latency = 0; 2251 2252 delay = _bundle_use.full_latency(relative_latency, node_usage); 2253 2254 // Does not fit in this bundle, start a new one 2255 if (delay > 0) { 2256 step(delay); 2257 2258 #ifndef PRODUCT 2259 if (_cfg->C->trace_opto_output()) 2260 tty->print("# *** STEP(%d) ***\n", delay); 2261 #endif 2262 } 2263 } 2264 2265 // If this was placed in the delay slot, ignore it 2266 if (n != _unconditional_delay_slot) { 2267 2268 if (delay == 0) { 2269 if (node_pipeline->hasMultipleBundles()) { 2270 #ifndef PRODUCT 2271 if (_cfg->C->trace_opto_output()) 2272 tty->print("# *** STEP(multiple instructions) ***\n"); 2273 #endif 2274 step(1); 2275 } 2276 2277 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2278 #ifndef PRODUCT 2279 if (_cfg->C->trace_opto_output()) 2280 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2281 instruction_count + _bundle_instr_count, 2282 Pipeline::_max_instrs_per_cycle); 2283 #endif 2284 step(1); 2285 } 2286 } 2287 2288 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2289 _bundle_instr_count++; 2290 2291 // Set the node's latency 2292 _current_latency[n->_idx] = _bundle_cycle_number; 2293 2294 // Now merge the functional unit information 2295 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2296 _bundle_use.add_usage(node_usage); 2297 2298 // Increment the number of instructions in this bundle 2299 _bundle_instr_count += instruction_count; 2300 2301 // Remember this node for later 2302 if (n->is_Mach()) 2303 _next_node = n; 2304 } 2305 2306 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2307 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2308 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2309 // into the block. All other scheduled nodes get put in the schedule here. 2310 int op = n->Opcode(); 2311 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2312 (op != Op_Node && // Not an unused antidepedence node and 2313 // not an unallocated boxlock 2314 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2315 2316 // Push any trailing projections 2317 if( bb->get_node(bb->number_of_nodes()-1) != n ) { 2318 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2319 Node *foi = n->fast_out(i); 2320 if( foi->is_Proj() ) 2321 _scheduled.push(foi); 2322 } 2323 } 2324 2325 // Put the instruction in the schedule list 2326 _scheduled.push(n); 2327 } 2328 2329 #ifndef PRODUCT 2330 if (_cfg->C->trace_opto_output()) 2331 dump_available(); 2332 #endif 2333 2334 // Walk all the definitions, decrementing use counts, and 2335 // if a definition has a 0 use count, place it in the available list. 2336 DecrementUseCounts(n,bb); 2337 } 2338 2339 // This method sets the use count within a basic block. We will ignore all 2340 // uses outside the current basic block. As we are doing a backwards walk, 2341 // any node we reach that has a use count of 0 may be scheduled. This also 2342 // avoids the problem of cyclic references from phi nodes, as long as phi 2343 // nodes are at the front of the basic block. This method also initializes 2344 // the available list to the set of instructions that have no uses within this 2345 // basic block. 2346 void Scheduling::ComputeUseCount(const Block *bb) { 2347 #ifndef PRODUCT 2348 if (_cfg->C->trace_opto_output()) 2349 tty->print("# -> ComputeUseCount\n"); 2350 #endif 2351 2352 // Clear the list of available and scheduled instructions, just in case 2353 _available.clear(); 2354 _scheduled.clear(); 2355 2356 // No delay slot specified 2357 _unconditional_delay_slot = NULL; 2358 2359 #ifdef ASSERT 2360 for( uint i=0; i < bb->number_of_nodes(); i++ ) 2361 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" ); 2362 #endif 2363 2364 // Force the _uses count to never go to zero for unscheduable pieces 2365 // of the block 2366 for( uint k = 0; k < _bb_start; k++ ) 2367 _uses[bb->get_node(k)->_idx] = 1; 2368 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ ) 2369 _uses[bb->get_node(l)->_idx] = 1; 2370 2371 // Iterate backwards over the instructions in the block. Don't count the 2372 // branch projections at end or the block header instructions. 2373 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2374 Node *n = bb->get_node(j); 2375 if( n->is_Proj() ) continue; // Projections handled another way 2376 2377 // Account for all uses 2378 for ( uint k = 0; k < n->len(); k++ ) { 2379 Node *inp = n->in(k); 2380 if (!inp) continue; 2381 assert(inp != n, "no cycles allowed" ); 2382 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use? 2383 if (inp->is_Proj()) { // Skip through Proj's 2384 inp = inp->in(0); 2385 } 2386 ++_uses[inp->_idx]; // Count 1 block-local use 2387 } 2388 } 2389 2390 // If this instruction has a 0 use count, then it is available 2391 if (!_uses[n->_idx]) { 2392 _current_latency[n->_idx] = _bundle_cycle_number; 2393 AddNodeToAvailableList(n); 2394 } 2395 2396 #ifndef PRODUCT 2397 if (_cfg->C->trace_opto_output()) { 2398 tty->print("# uses: %3d: ", _uses[n->_idx]); 2399 n->dump(); 2400 } 2401 #endif 2402 } 2403 2404 #ifndef PRODUCT 2405 if (_cfg->C->trace_opto_output()) 2406 tty->print("# <- ComputeUseCount\n"); 2407 #endif 2408 } 2409 2410 // This routine performs scheduling on each basic block in reverse order, 2411 // using instruction latencies and taking into account function unit 2412 // availability. 2413 void Scheduling::DoScheduling() { 2414 #ifndef PRODUCT 2415 if (_cfg->C->trace_opto_output()) 2416 tty->print("# -> DoScheduling\n"); 2417 #endif 2418 2419 Block *succ_bb = NULL; 2420 Block *bb; 2421 2422 // Walk over all the basic blocks in reverse order 2423 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) { 2424 bb = _cfg->get_block(i); 2425 2426 #ifndef PRODUCT 2427 if (_cfg->C->trace_opto_output()) { 2428 tty->print("# Schedule BB#%03d (initial)\n", i); 2429 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2430 bb->get_node(j)->dump(); 2431 } 2432 } 2433 #endif 2434 2435 // On the head node, skip processing 2436 if (bb == _cfg->get_root_block()) { 2437 continue; 2438 } 2439 2440 // Skip empty, connector blocks 2441 if (bb->is_connector()) 2442 continue; 2443 2444 // If the following block is not the sole successor of 2445 // this one, then reset the pipeline information 2446 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2447 #ifndef PRODUCT 2448 if (_cfg->C->trace_opto_output()) { 2449 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2450 _next_node->_idx, _bundle_instr_count); 2451 } 2452 #endif 2453 step_and_clear(); 2454 } 2455 2456 // Leave untouched the starting instruction, any Phis, a CreateEx node 2457 // or Top. bb->get_node(_bb_start) is the first schedulable instruction. 2458 _bb_end = bb->number_of_nodes()-1; 2459 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2460 Node *n = bb->get_node(_bb_start); 2461 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2462 // Also, MachIdealNodes do not get scheduled 2463 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2464 MachNode *mach = n->as_Mach(); 2465 int iop = mach->ideal_Opcode(); 2466 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2467 if( iop == Op_Con ) continue; // Do not schedule Top 2468 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2469 mach->pipeline() == MachNode::pipeline_class() && 2470 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc 2471 continue; 2472 break; // Funny loop structure to be sure... 2473 } 2474 // Compute last "interesting" instruction in block - last instruction we 2475 // might schedule. _bb_end points just after last schedulable inst. We 2476 // normally schedule conditional branches (despite them being forced last 2477 // in the block), because they have delay slots we can fill. Calls all 2478 // have their delay slots filled in the template expansions, so we don't 2479 // bother scheduling them. 2480 Node *last = bb->get_node(_bb_end); 2481 // Ignore trailing NOPs. 2482 while (_bb_end > 0 && last->is_Mach() && 2483 last->as_Mach()->ideal_Opcode() == Op_Con) { 2484 last = bb->get_node(--_bb_end); 2485 } 2486 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2487 if( last->is_Catch() || 2488 // Exclude unreachable path case when Halt node is in a separate block. 2489 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2490 // There must be a prior call. Skip it. 2491 while( !bb->get_node(--_bb_end)->is_MachCall() ) { 2492 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" ); 2493 } 2494 } else if( last->is_MachNullCheck() ) { 2495 // Backup so the last null-checked memory instruction is 2496 // outside the schedulable range. Skip over the nullcheck, 2497 // projection, and the memory nodes. 2498 Node *mem = last->in(1); 2499 do { 2500 _bb_end--; 2501 } while (mem != bb->get_node(_bb_end)); 2502 } else { 2503 // Set _bb_end to point after last schedulable inst. 2504 _bb_end++; 2505 } 2506 2507 assert( _bb_start <= _bb_end, "inverted block ends" ); 2508 2509 // Compute the register antidependencies for the basic block 2510 ComputeRegisterAntidependencies(bb); 2511 if (_cfg->C->failing()) return; // too many D-U pinch points 2512 2513 // Compute intra-bb latencies for the nodes 2514 ComputeLocalLatenciesForward(bb); 2515 2516 // Compute the usage within the block, and set the list of all nodes 2517 // in the block that have no uses within the block. 2518 ComputeUseCount(bb); 2519 2520 // Schedule the remaining instructions in the block 2521 while ( _available.size() > 0 ) { 2522 Node *n = ChooseNodeToBundle(); 2523 guarantee(n != NULL, "no nodes available"); 2524 AddNodeToBundle(n,bb); 2525 } 2526 2527 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2528 #ifdef ASSERT 2529 for( uint l = _bb_start; l < _bb_end; l++ ) { 2530 Node *n = bb->get_node(l); 2531 uint m; 2532 for( m = 0; m < _bb_end-_bb_start; m++ ) 2533 if( _scheduled[m] == n ) 2534 break; 2535 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2536 } 2537 #endif 2538 2539 // Now copy the instructions (in reverse order) back to the block 2540 for ( uint k = _bb_start; k < _bb_end; k++ ) 2541 bb->map_node(_scheduled[_bb_end-k-1], k); 2542 2543 #ifndef PRODUCT 2544 if (_cfg->C->trace_opto_output()) { 2545 tty->print("# Schedule BB#%03d (final)\n", i); 2546 uint current = 0; 2547 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2548 Node *n = bb->get_node(j); 2549 if( valid_bundle_info(n) ) { 2550 Bundle *bundle = node_bundling(n); 2551 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2552 tty->print("*** Bundle: "); 2553 bundle->dump(); 2554 } 2555 n->dump(); 2556 } 2557 } 2558 } 2559 #endif 2560 #ifdef ASSERT 2561 verify_good_schedule(bb,"after block local scheduling"); 2562 #endif 2563 } 2564 2565 #ifndef PRODUCT 2566 if (_cfg->C->trace_opto_output()) 2567 tty->print("# <- DoScheduling\n"); 2568 #endif 2569 2570 // Record final node-bundling array location 2571 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2572 2573 } // end DoScheduling 2574 2575 // Verify that no live-range used in the block is killed in the block by a 2576 // wrong DEF. This doesn't verify live-ranges that span blocks. 2577 2578 // Check for edge existence. Used to avoid adding redundant precedence edges. 2579 static bool edge_from_to( Node *from, Node *to ) { 2580 for( uint i=0; i<from->len(); i++ ) 2581 if( from->in(i) == to ) 2582 return true; 2583 return false; 2584 } 2585 2586 #ifdef ASSERT 2587 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2588 // Check for bad kills 2589 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2590 Node *prior_use = _reg_node[def]; 2591 if( prior_use && !edge_from_to(prior_use,n) ) { 2592 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2593 n->dump(); 2594 tty->print_cr("..."); 2595 prior_use->dump(); 2596 assert(edge_from_to(prior_use,n),msg); 2597 } 2598 _reg_node.map(def,NULL); // Kill live USEs 2599 } 2600 } 2601 2602 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2603 2604 // Zap to something reasonable for the verify code 2605 _reg_node.clear(); 2606 2607 // Walk over the block backwards. Check to make sure each DEF doesn't 2608 // kill a live value (other than the one it's supposed to). Add each 2609 // USE to the live set. 2610 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) { 2611 Node *n = b->get_node(i); 2612 int n_op = n->Opcode(); 2613 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2614 // Fat-proj kills a slew of registers 2615 RegMask rm = n->out_RegMask();// Make local copy 2616 while( rm.is_NotEmpty() ) { 2617 OptoReg::Name kill = rm.find_first_elem(); 2618 rm.Remove(kill); 2619 verify_do_def( n, kill, msg ); 2620 } 2621 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2622 // Get DEF'd registers the normal way 2623 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2624 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2625 } 2626 2627 // Now make all USEs live 2628 for( uint i=1; i<n->req(); i++ ) { 2629 Node *def = n->in(i); 2630 assert(def != 0, "input edge required"); 2631 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2632 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2633 if( OptoReg::is_valid(reg_lo) ) { 2634 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg); 2635 _reg_node.map(reg_lo,n); 2636 } 2637 if( OptoReg::is_valid(reg_hi) ) { 2638 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg); 2639 _reg_node.map(reg_hi,n); 2640 } 2641 } 2642 2643 } 2644 2645 // Zap to something reasonable for the Antidependence code 2646 _reg_node.clear(); 2647 } 2648 #endif 2649 2650 // Conditionally add precedence edges. Avoid putting edges on Projs. 2651 static void add_prec_edge_from_to( Node *from, Node *to ) { 2652 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2653 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2654 from = from->in(0); 2655 } 2656 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2657 !edge_from_to( from, to ) ) // Avoid duplicate edge 2658 from->add_prec(to); 2659 } 2660 2661 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2662 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2663 return; 2664 2665 Node *pinch = _reg_node[def_reg]; // Get pinch point 2666 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet? 2667 is_def ) { // Check for a true def (not a kill) 2668 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2669 return; 2670 } 2671 2672 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2673 debug_only( def = (Node*)0xdeadbeef; ) 2674 2675 // After some number of kills there _may_ be a later def 2676 Node *later_def = NULL; 2677 2678 // Finding a kill requires a real pinch-point. 2679 // Check for not already having a pinch-point. 2680 // Pinch points are Op_Node's. 2681 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2682 later_def = pinch; // Must be def/kill as optimistic pinch-point 2683 if ( _pinch_free_list.size() > 0) { 2684 pinch = _pinch_free_list.pop(); 2685 } else { 2686 pinch = new (_cfg->C) Node(1); // Pinch point to-be 2687 } 2688 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2689 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2690 return; 2691 } 2692 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init) 2693 _reg_node.map(def_reg,pinch); // Record pinch-point 2694 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2695 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2696 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2697 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2698 later_def = NULL; // and no later def 2699 } 2700 pinch->set_req(0,later_def); // Hook later def so we can find it 2701 } else { // Else have valid pinch point 2702 if( pinch->in(0) ) // If there is a later-def 2703 later_def = pinch->in(0); // Get it 2704 } 2705 2706 // Add output-dependence edge from later def to kill 2707 if( later_def ) // If there is some original def 2708 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2709 2710 // See if current kill is also a use, and so is forced to be the pinch-point. 2711 if( pinch->Opcode() == Op_Node ) { 2712 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2713 for( uint i=1; i<uses->req(); i++ ) { 2714 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2715 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2716 // Yes, found a use/kill pinch-point 2717 pinch->set_req(0,NULL); // 2718 pinch->replace_by(kill); // Move anti-dep edges up 2719 pinch = kill; 2720 _reg_node.map(def_reg,pinch); 2721 return; 2722 } 2723 } 2724 } 2725 2726 // Add edge from kill to pinch-point 2727 add_prec_edge_from_to(kill,pinch); 2728 } 2729 2730 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2731 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2732 return; 2733 Node *pinch = _reg_node[use_reg]; // Get pinch point 2734 // Check for no later def_reg/kill in block 2735 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b && 2736 // Use has to be block-local as well 2737 _cfg->get_block_for_node(use) == b) { 2738 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2739 pinch->req() == 1 ) { // pinch not yet in block? 2740 pinch->del_req(0); // yank pointer to later-def, also set flag 2741 // Insert the pinch-point in the block just after the last use 2742 b->insert_node(pinch, b->find_node(use) + 1); 2743 _bb_end++; // Increase size scheduled region in block 2744 } 2745 2746 add_prec_edge_from_to(pinch,use); 2747 } 2748 } 2749 2750 // We insert antidependences between the reads and following write of 2751 // allocated registers to prevent illegal code motion. Hopefully, the 2752 // number of added references should be fairly small, especially as we 2753 // are only adding references within the current basic block. 2754 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2755 2756 #ifdef ASSERT 2757 verify_good_schedule(b,"before block local scheduling"); 2758 #endif 2759 2760 // A valid schedule, for each register independently, is an endless cycle 2761 // of: a def, then some uses (connected to the def by true dependencies), 2762 // then some kills (defs with no uses), finally the cycle repeats with a new 2763 // def. The uses are allowed to float relative to each other, as are the 2764 // kills. No use is allowed to slide past a kill (or def). This requires 2765 // antidependencies between all uses of a single def and all kills that 2766 // follow, up to the next def. More edges are redundant, because later defs 2767 // & kills are already serialized with true or antidependencies. To keep 2768 // the edge count down, we add a 'pinch point' node if there's more than 2769 // one use or more than one kill/def. 2770 2771 // We add dependencies in one bottom-up pass. 2772 2773 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2774 2775 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2776 // register. If not, we record the DEF/KILL in _reg_node, the 2777 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2778 // "pinch point", a new Node that's in the graph but not in the block. 2779 // We put edges from the prior and current DEF/KILLs to the pinch point. 2780 // We put the pinch point in _reg_node. If there's already a pinch point 2781 // we merely add an edge from the current DEF/KILL to the pinch point. 2782 2783 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2784 // put an edge from the pinch point to the USE. 2785 2786 // To be expedient, the _reg_node array is pre-allocated for the whole 2787 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2788 // or a valid def/kill/pinch-point, or a leftover node from some prior 2789 // block. Leftover node from some prior block is treated like a NULL (no 2790 // prior def, so no anti-dependence needed). Valid def is distinguished by 2791 // it being in the current block. 2792 bool fat_proj_seen = false; 2793 uint last_safept = _bb_end-1; 2794 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL; 2795 Node* last_safept_node = end_node; 2796 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2797 Node *n = b->get_node(i); 2798 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2799 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2800 // Fat-proj kills a slew of registers 2801 // This can add edges to 'n' and obscure whether or not it was a def, 2802 // hence the is_def flag. 2803 fat_proj_seen = true; 2804 RegMask rm = n->out_RegMask();// Make local copy 2805 while( rm.is_NotEmpty() ) { 2806 OptoReg::Name kill = rm.find_first_elem(); 2807 rm.Remove(kill); 2808 anti_do_def( b, n, kill, is_def ); 2809 } 2810 } else { 2811 // Get DEF'd registers the normal way 2812 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2813 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2814 } 2815 2816 // Kill projections on a branch should appear to occur on the 2817 // branch, not afterwards, so grab the masks from the projections 2818 // and process them. 2819 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) { 2820 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2821 Node* use = n->fast_out(i); 2822 if (use->is_Proj()) { 2823 RegMask rm = use->out_RegMask();// Make local copy 2824 while( rm.is_NotEmpty() ) { 2825 OptoReg::Name kill = rm.find_first_elem(); 2826 rm.Remove(kill); 2827 anti_do_def( b, n, kill, false ); 2828 } 2829 } 2830 } 2831 } 2832 2833 // Check each register used by this instruction for a following DEF/KILL 2834 // that must occur afterward and requires an anti-dependence edge. 2835 for( uint j=0; j<n->req(); j++ ) { 2836 Node *def = n->in(j); 2837 if( def ) { 2838 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2839 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2840 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2841 } 2842 } 2843 // Do not allow defs of new derived values to float above GC 2844 // points unless the base is definitely available at the GC point. 2845 2846 Node *m = b->get_node(i); 2847 2848 // Add precedence edge from following safepoint to use of derived pointer 2849 if( last_safept_node != end_node && 2850 m != last_safept_node) { 2851 for (uint k = 1; k < m->req(); k++) { 2852 const Type *t = m->in(k)->bottom_type(); 2853 if( t->isa_oop_ptr() && 2854 t->is_ptr()->offset() != 0 ) { 2855 last_safept_node->add_prec( m ); 2856 break; 2857 } 2858 } 2859 } 2860 2861 if( n->jvms() ) { // Precedence edge from derived to safept 2862 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2863 if( b->get_node(last_safept) != last_safept_node ) { 2864 last_safept = b->find_node(last_safept_node); 2865 } 2866 for( uint j=last_safept; j > i; j-- ) { 2867 Node *mach = b->get_node(j); 2868 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2869 mach->add_prec( n ); 2870 } 2871 last_safept = i; 2872 last_safept_node = m; 2873 } 2874 } 2875 2876 if (fat_proj_seen) { 2877 // Garbage collect pinch nodes that were not consumed. 2878 // They are usually created by a fat kill MachProj for a call. 2879 garbage_collect_pinch_nodes(); 2880 } 2881 } 2882 2883 // Garbage collect pinch nodes for reuse by other blocks. 2884 // 2885 // The block scheduler's insertion of anti-dependence 2886 // edges creates many pinch nodes when the block contains 2887 // 2 or more Calls. A pinch node is used to prevent a 2888 // combinatorial explosion of edges. If a set of kills for a 2889 // register is anti-dependent on a set of uses (or defs), rather 2890 // than adding an edge in the graph between each pair of kill 2891 // and use (or def), a pinch is inserted between them: 2892 // 2893 // use1 use2 use3 2894 // \ | / 2895 // \ | / 2896 // pinch 2897 // / | \ 2898 // / | \ 2899 // kill1 kill2 kill3 2900 // 2901 // One pinch node is created per register killed when 2902 // the second call is encountered during a backwards pass 2903 // over the block. Most of these pinch nodes are never 2904 // wired into the graph because the register is never 2905 // used or def'ed in the block. 2906 // 2907 void Scheduling::garbage_collect_pinch_nodes() { 2908 #ifndef PRODUCT 2909 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2910 #endif 2911 int trace_cnt = 0; 2912 for (uint k = 0; k < _reg_node.Size(); k++) { 2913 Node* pinch = _reg_node[k]; 2914 if ((pinch != NULL) && pinch->Opcode() == Op_Node && 2915 // no predecence input edges 2916 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2917 cleanup_pinch(pinch); 2918 _pinch_free_list.push(pinch); 2919 _reg_node.map(k, NULL); 2920 #ifndef PRODUCT 2921 if (_cfg->C->trace_opto_output()) { 2922 trace_cnt++; 2923 if (trace_cnt > 40) { 2924 tty->print("\n"); 2925 trace_cnt = 0; 2926 } 2927 tty->print(" %d", pinch->_idx); 2928 } 2929 #endif 2930 } 2931 } 2932 #ifndef PRODUCT 2933 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2934 #endif 2935 } 2936 2937 // Clean up a pinch node for reuse. 2938 void Scheduling::cleanup_pinch( Node *pinch ) { 2939 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2940 2941 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2942 Node* use = pinch->last_out(i); 2943 uint uses_found = 0; 2944 for (uint j = use->req(); j < use->len(); j++) { 2945 if (use->in(j) == pinch) { 2946 use->rm_prec(j); 2947 uses_found++; 2948 } 2949 } 2950 assert(uses_found > 0, "must be a precedence edge"); 2951 i -= uses_found; // we deleted 1 or more copies of this edge 2952 } 2953 // May have a later_def entry 2954 pinch->set_req(0, NULL); 2955 } 2956 2957 #ifndef PRODUCT 2958 2959 void Scheduling::dump_available() const { 2960 tty->print("#Availist "); 2961 for (uint i = 0; i < _available.size(); i++) 2962 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2963 tty->cr(); 2964 } 2965 2966 // Print Scheduling Statistics 2967 void Scheduling::print_statistics() { 2968 // Print the size added by nops for bundling 2969 tty->print("Nops added %d bytes to total of %d bytes", 2970 _total_nop_size, _total_method_size); 2971 if (_total_method_size > 0) 2972 tty->print(", for %.2f%%", 2973 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2974 tty->print("\n"); 2975 2976 // Print the number of branch shadows filled 2977 if (Pipeline::_branch_has_delay_slot) { 2978 tty->print("Of %d branches, %d had unconditional delay slots filled", 2979 _total_branches, _total_unconditional_delays); 2980 if (_total_branches > 0) 2981 tty->print(", for %.2f%%", 2982 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2983 tty->print("\n"); 2984 } 2985 2986 uint total_instructions = 0, total_bundles = 0; 2987 2988 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2989 uint bundle_count = _total_instructions_per_bundle[i]; 2990 total_instructions += bundle_count * i; 2991 total_bundles += bundle_count; 2992 } 2993 2994 if (total_bundles > 0) 2995 tty->print("Average ILP (excluding nops) is %.2f\n", 2996 ((double)total_instructions) / ((double)total_bundles)); 2997 } 2998 #endif